Two announcements this weekend were both about insurance. Microsoft’s chief executive said out loud what security teams have assumed for a year, that a model you cannot audit is a model you should treat as already breached, and asked for a brake a human can pull. Meanwhile the competitive answer to that risk arrived from the other direction: Nvidia, already the largest single investor in the open-weight lab Reflection AI, is discussing buying it outright, which would put the industry’s most-watched open model inside the company that supplies the chips it was trained on. The through-line running under both is money, and this issue is unusually full of numbers: a compute contract billed monthly, a valuation in the tens of billions, a commission counting lost healthy life-years in the billions, and a developer offering a rural township ten thousand dollars per household.

⏰ OpenRouter Limited-Time Free Models

This section needs two passes, because the collector caps its list at ten rows and those rows are an expiration_date list rather than a free list. This cycle the collector returned ten dated rows of which exactly one is genuinely free; sweeping every published sunset across the full model list returns 18 dated endpoints, unchanged from yesterday; filtering to :free IDs separates the permanent shelf from the promotions.

Model Context Price (in / out) Sunset
poolside/laguna-s-2.1:free 262,144 free 2026-10-31
poolside/laguna-xs-2.1:free 262,144 free 2026-10-31
dots-studio/dots-3-note-preview:free 512,000 free 2026-12-31
google/gemini-2.5-flash-lite 1,048,576 $0.10 / $0.40 2026-10-20
google/gemini-2.5-flash 1,048,576 $0.30 / $2.50 2026-10-20
google/gemini-2.5-pro 1,048,576 $1.25 / $10.00 2026-10-20
qwen/qwen3.5-9b 262,144 $0.10 / $0.15 2026-10-21
xiaomi/mimo-v2.5 1,050,000 $0.14 / $0.28 2026-10-21
xiaomi/mimo-v2.5-pro 1,050,000 $0.435 / $0.87 2026-10-21
poolside/laguna-s-2.1 1,048,576 $0.09 / $0.18 2026-10-31

The calendar is now stable rather than shrinking. Yesterday’s sweep found 18 dated endpoints after the 15-endpoint Qwen deprecation batch cleared on schedule; today it finds 18 again, so nothing lapsed in the last 24 hours and nothing new was scheduled. What remains clusters into five batches: five Gemini 2.5 endpoints on 2026-10-20, three on 2026-10-21 (qwen3.5-9b and the two Xiaomi MiMo-V2.5 variants), four Poolside Laguna endpoints on 2026-10-31, three ByteDance Seed models on 2026-11-11, and three on 2026-12-31 (the Z.ai GLM pair plus the free Dots preview).

Three limited-time free endpoints are live, and two of them — the Poolside Laguna S and XS variants — are only visible in the full sweep, since the collector’s ten-row cap caught just one. The permanent :free shelf held at 15 endpoints with 12 of them undated, unchanged for a second day.

🤗 Trending on Hugging Face

The collector’s new-uploads list was zero-download test spam for the 37th consecutive cycle, dominated by single-purpose training checkpoints such as yurunyyr/asyncrl-sync_mb8_kltheta_vs_old_b0p05_truncneg1_replay_rho0p125_ev128-gs300. This section therefore uses Hugging Face’s official trending feed, which ranks by recent popularity rather than upload time.

Model Downloads Likes Task
Qwen/Qwen3.8-27B 6,843,108 17,474 image-text-to-text
abenzerps/Qwen-Image-2.1-Uncensored-GGUF 2,188,658 3,930 text-to-image
Lightricks/LTX-2.5 1,726,683 7,232 image-to-video
XiaomiMiMo/MiMo-V2.6-RL-oss 113,898 908 text-generation
google/embeddinggemma-2 98,829 1,601 feature-extraction
LightwheelAI/EgoDemo 61,148 187 robotics video
Venastine-Research/Xing4.0-29B-A4B-GGUF 48,389 688 text-generation
jialinyyzz/humanizer 37,647 1,008 text-generation
Cloudflare/clef 14,351 2,012 image-text-to-text
Aleph-Alpha/Kolibri-1 12,550 879 text-generation

The feed’s biggest single-day jump belongs to google/embeddinggemma-2, which more than doubled from 45,605 downloads to 98,829 while its likes moved only 1,425 to 1,601 — the signature of a pipeline pulling an embedding model rather than a crowd discovering it. Two new entrants are worth naming. Aleph-Alpha/Kolibri-1 is a fresh release from the German lab, arriving with 879 likes on its first day, and jialinyyzz/humanizer turned up at 37,647 downloads and 1,008 likes with no announcement attached. The Jev thread, meanwhile, has stopped producing models and started producing plumbing: multimodalart/jev-decision-index is still on the feed at 583 likes with no recorded downloads, tardellirs/model-pulse sits at 182 likes, and yesterday’s leader LocalLLaMA/typed-decisions fell off the feed entirely after reaching 36,648 downloads.

⭐ GitHub Trending: AI Edition

All five of the collector’s entries were out of window: mhtsec/ARTEX was created 2026-10-08, ThariqS/ai-newtab and franzenzenhofer/big-arrow-on-the-screen on 2026-10-08, LosaLosSantos/aurelio-finance on 2026-10-06, and rociiu/talorys — yesterday’s top entry — on 2026-10-10 at 09:46 UTC, before the previous issue shipped at 13:08. The repositories below were verified individually against the GitHub API and created inside the last 24 hours.

Repository Language Stars What it is
Sam-Hui-dot/mcd-budget-meal Python 70 A budget-meal planner built on the official McDonald’s China MCP
csadigital/ai-writer JavaScript 34 Multilingual product copy generated from images
artouc/stop-describing-ui Markdown 22 An agent skill that strips explanatory copy out of interfaces
hongboduan/awesome-embodied-agents Python 19 A reading list for embodied agents
Mint-Security/claude-code-plugins TypeScript 16 Security plugins for Claude Code, from a security vendor
final-bench/quantum-bench-harness Python 15 A frozen evaluation harness for a quantum leaderboard
2289698809/mcfit-agent Python 12 A calorie counter and diet planner on the same McDonald’s MCP
scs0209/laya-guard Python 3 An on-device guardrail that judges every tool call before it runs

The pool was genuinely thin — the highest-starred repository created in the window was not AI at all, a semiconductor manufacturing primer at 96 stars, and the top AI entry reached only 70. That 70 belongs to a contest. McDonald’s China runs an official MCP platform and a 2026 developer innovation challenge around it, and this is its third day of output: mcd-budget-meal estimates a meal against a budget, mcfit-agent does the same against calories and sodium, and ftxcgwf/mcd-solver frames it as constrained optimisation. Two manufactured clusters were excluded by shape rather than by choice: a dozen Windows game-utility repositories created inside eight seconds at 12:32 UTC at exactly 30–31 stars, and the alphaparkinc / Alpha-Park pair publishing dozens of identically formatted genpark-* agent skills at exactly 7 stars apiece.

🏭 Nvidia Moves on the Lab That Trains on Its Chips

Nvidia is in early talks to either buy Reflection AI outright or deepen its stake in it, the Financial Times reported on Saturday, and Bloomberg and Reuters followed on Sunday with the detail that the discussion also covers a computing deal. Nvidia has already put 800 million dollars into the company. The startup, founded in 2024 by the former DeepMind researchers Misha Laskin and Ioannis Antonoglou, was reported as a 25 billion dollar business, and its first open-weight model, Beam, was unveiled on 2026-10-05 and trained on Nvidia GB300 GPUs. The options on the table are a full acquisition, additional financing, or an acqui-hire in which Nvidia takes the staff and licenses the technology. Talks are early and could still collapse.

Read the shape of the deal rather than the headline. The most-watched independent open-weight lab of the moment runs on the seller’s hardware, and the seller is the buyer. If the acqui-hire version wins, the model weights stay open while the people who built them stop being independent of the chip supply they depend on. This is the same vertical logic that has been visible all year — Anthropic renting from SpaceX, Google guaranteeing leases — applied one layer down, at the lab rather than the datacenter.

🛑 Assume the Model Is Compromised, Then Build It a Brake

Satya Nadella published a short position on X on Saturday that the industry has been circling for months: companies should treat powerful models as potential insider threats, assume they may already be compromised, and contain them from the start. The Microsoft chief called for systems that include containment, independent controls and an emergency brake a human can pull mid-task, alongside continuous testing and transparent monitoring.

The concrete detail underneath the philosophy came from a CNBC report on the same day, which found that the third-party evaluators being asked to carry this load — Model Evaluation and Threat Research, Apollo Research and Transluce — are mostly nonprofits with unsettled funding. Anthropic’s pledge to embed independent evaluators, endorsed within hours by OpenAI, supported by the White House and by most large US technology companies, left three questions unanswered: who pays them, how much access they get, and who they report to. Suresh Venkatasubramanian of Brown University put the problem plainly to CNBC: an ecosystem needs a viable business model, and right now it does not have one. The company writing the rules is the company being evaluated.

⚖️ Anthropic’s Political Fixer, a Monthly Bill, and a Two Trillion Dollar Listing

A Wall Street Journal report published over the weekend reconstructs how Anthropic co-founder and chief compute officer Tom Brown used his Republican relationships to resolve a June standoff in which the White House restricted access to two Anthropic models on security grounds, negotiating with the Commerce and Defence departments and the White House directly while assuring officials that the company had upgraded its safeguards. The same reporting credits Brown with repairing relations with Elon Musk and brokering the compute agreement that has Anthropic paying SpaceX 1.25 billion dollars per month.

The monthly figure is the part that is easy to read past. The contract runs through May 2029 and has been valued at roughly 45 billion dollars over its term, with either party able to terminate on 90 days notice, and it buys more than 300 megawatts and over 220,000 Nvidia GPUs. SpaceX’s own filing describes it as a way to monetise unused capacity, which is a rare thing for an AI supplier to admit out loud: the seller is renting out a surplus, and the buyer is booking it as a strategic necessity. Anthropic is pursuing an initial public offering that would value it above two trillion dollars.

☢️ The Lancet Put Malicious AI and Nuclear War in a Category of Two

A new Lancet Commission on global health threats identified 17 interconnected risks that could cost humanity tens of billions of healthy life-years by 2100, and named exactly two of them as capable of ending the species: nuclear conflict and the malicious use of artificial intelligence. The commission, 35 leaders drawing on the Institute for Health Metrics and Evaluation’s Global Burden of Disease data, puts the estimate for malicious AI at about 10.4 billion healthy life-years lost by 2100, while stressing that the projection is theoretical and that forecasters disagree widely. Its co-chair, Christopher Murray, framed the finding around agency rather than doom: humanity retains the ability to act, and delay shifts the cost forward.

It is worth being precise about what the report claims and what it does not. The commission does not say AI is likely to end humanity; it says AI and nuclear war are the only two threats on a list of 17 whose worst case is extinction, and that obesity, high blood pressure and smoking — the threats with far stronger empirical support — sit lower on the severity scale even though they will certainly kill far more people. That inversion is the interesting part. The threats we can measure are the ones we have decided not to fund.

🦾 The Attack Chain Reached the Robotic Arm

Researchers at Booz Allen Hamilton ran eight scenarios in which autonomous AI systems were pointed at operational technology — the industrial equipment that controls water, power and manufacturing — and the models achieved the objective in all eight, according to a report covered by The Register. The agents located weak links in control systems and network configurations, and in one scenario found and moved a robotic arm in minutes. The researchers also report agents making operator screens on OT devices display false readings, which is the detail that turns a breach into a physical event, because it removes the human who would otherwise notice.

Kyle Miller, the firm’s vice president of infrastructure cybersecurity, told The Register that agents are operating with a speed, persistence and engineering precision that will outpace organisations without foundational OT security practices. The uncomfortable conclusion is about expertise rather than capability: an attacker no longer needs to understand industrial systems, because the model does. Defenders, meanwhile, may have minutes.

🏗 Data Centers Became the Bipartisan Common Enemy

Fortune’s weekend feature on Hazle Township, Pennsylvania is a case study in how quickly AI infrastructure turned into a local political problem. NorthPoint Development offered the township’s 4,500 residents a potential 10,000 dollars per household in exchange for permission to build a 1,300-acre, 15-building data center campus as part of a 165 million dollar investment; roughly 30 nearby properties then faced eminent domain threats from a utility’s high-voltage power line serving the same project. Locals filed about 20 formal protests, and right-to-know requests established that the end user is Amazon — a fact absent from every piece of NorthPoint and Project Hazelnut branding. Amazon says it is evaluating multiple sites and has no definitive agreement.

The numbers around that fight are the story. About 70 percent of Americans oppose a data center being built in their area, including 63 percent of Republicans. At least 75 projects worth more than 130 billion dollars were delayed or cancelled in the first three months of 2026, and active opposition groups grew from 396 to 833 over the same period. Pennsylvania’s governor has already eliminated the fast-track permitting order that launched the project. The pricing angle is arriving in court at the same time: a class action filed in Illinois alleges that McDonald’s AI-assisted pricing tool lets competing franchisees share non-public sales data, which the plaintiff calls algorithmic price-fixing, and which McDonald’s denies on the grounds that franchisees, not software, set menu prices. A Reuters/Ipsos poll released on 2026-10-07 found 84 percent of US voters see AI as a threat to American workers and 56 percent support limits on new data center construction. The buildout now has a cost curve that is political before it is technical.

📌 Quick Hits

  • Nicolas Cage refused to sign an AI waiver for the Amazon series Spider-Noir and said he probably will not work with Prime Video again, telling reporters he is not an AI-friendly actor.
  • The Register reports HPE’s networking chief predicting AI will resolve all trouble tickets without human involvement within two years.
  • A University of Limerick study found a female-presenting AI agent was paid 10.25 percent less than a male-presenting one for identical work, using the same underlying model, with 189 participants in a VR office dividing real money.
  • Wired’s reporting, via Gizmodo, found three of the big five book publishers using ChatGPT and Claude for publicity copy, cover text and cover art without telling authors.
  • The Washington Post argues the US and China are unintentionally allied in the AI race, through the supply chains each needs from the other.
  • Prime Intellect rewrote Prime Agent in Rust, and a Berkeley study on AI use eroding persistence stayed on the Hacker News front page for a second day.
  • A 500-billion-token experiment in which AI agents decompiled a first-person shooter reached 111 points on Hacker News, ahead of most of the weekend’s vendor news.

💡 Key Trends

  1. The safety conversation has moved from principles to plumbing. Nadella’s emergency brake, Anthropic’s independent evaluators, METR and Transluce being asked to carry the load, and the question of who pays them are all the same problem stated at different levels. The White House has settled on voluntary self-policing, which means the enforcement layer has to be built by nonprofits whose funding is unsettled. Every party agrees the function is necessary; nobody has agreed on where the money comes from.

  2. Consolidation is moving up the stack, toward the labs themselves. Nvidia buying Reflection AI, a lab it already part-funded and whose model was trained on its own GB300 chips, would fold an independent research group into its supply chain at the same moment Anthropic is renting capacity from SpaceX on a monthly contract. The pattern across the sector is that independence is becoming a line item someone else controls.

  3. The data center backlash has become the most reliable political signal in AI. Seventy percent local opposition, 75 projects worth 130 billion dollars delayed or cancelled in a quarter, and a protest movement that doubled in size, all without a federal mandate in either direction. Pennsylvania is being watched less because of the project than because it is a purple state with a governor expected to run for president in 2028.

  4. Risk frameworks have stopped treating AI as a category of its own. The Lancet’s 17-threat list puts malicious AI beside nuclear war and then puts obesity, smoking and air pollution near the top of the same list — a reminder that a threat being speculative does not make it unimportant, and a threat being measurable does not make it funded.

Today’s Snapshot

  • New OpenRouter models: 0 since 2026-10-10 — the created-timestamp sweep found nothing newer than stepfun/step-5-preview from 2026-10-08, so the New OpenRouter Models section is omitted this issue.
  • Limited-time free models: 3 live (Poolside Laguna S and XS at 262,144 context, Dots3-Note Preview at 512,000), all previously dated; permanent :free shelf 15, of which 12 undated.
  • Expiration batches coming: five Gemini 2.5 endpoints on 2026-10-20, three on 2026-10-21, four Poolside Laguna on 2026-10-31, three ByteDance Seed on 2026-11-11, three on 2026-12-31 — 18 dated endpoints in total.
  • Hugging Face: 37th consecutive cycle of zero-download spam in the collector list, trending feed used instead; 10 models listed, led by Qwen3.8-27B at 6,843,108 downloads.
  • GitHub: 0 of 5 collector entries in window; 8 API-verified substitutes, top entry 70 stars, two manufactured clusters excluded.
  • OpenRouter pricing: unchanged headline rates; the free shelf is the only place a 1M-context model costs nothing, and it does not exist on that shelf today.